Built-in restrictions or filters that prevent a model from generating harmful, illegal, or unethical content.